Claude writes 90% of Claude Code's own codebase. Here's the 7-layer review process Anthropic built so merged PRs doubled while reverts fell 40%.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Claude writes about 90% of the code in Claude Code itself. Not a demo stat, not a cherry-picked sprint. That's Anthropic's own engineering team describing their day-to-day reality as of October 2026. And instead of drowning in slop, their merged PRs more than doubled between June and September while the revert rate fell by over 40%. The team wrote up exactly how they pulled that off, and it's one of the more useful process documents I've read this year.
Here's the first thing worth sitting with: more AI-written code didn't mean more human review time. It meant less. The team's stated goal is that most pull requests merge with at most one human in the loop, and that human can be the author. That only works if you've redesigned review from the ground up, not just bolted a bot onto your existing GitHub flow.
Anthropic's monorepo went from roughly 2,500 changes a day in May 2026 to more than 30,000 a day by September: a 12x jump in four months, driven almost entirely by Claude agents writing, testing, and shipping code on their own. That's not a typo. Git itself started buckling under the load, which is a separate (and wild) story about Anthropic building a replacement source-control system called Scion. But the review process is the part every team can actually copy, and it's what this post is about.
Two numbers tell you the review system is working, not just absorbing volume: internal PRs that pass in one review round went from 65% to 78%, and the average number of review rounds per PR dropped from 2.08 to 1.32. Fewer back-and-forths, more first-try merges, at 12x the volume.
It's seven layers, and only the last couple involve a human reading anything line by line.
CLAUDE.md and REVIEW.md conventions.stable channel.That last point matters more than it sounds. Every PR that caused a real incident becomes a permanent eval case, and every change to the review bot gets scored against that whole incident set before it ships. The review system itself gets code-reviewed, basically.
Because the team stopped treating "read every line" as the safety mechanism and moved trust to tests, agent-generated explainers, and automated analysis instead. Boris Cherny's line from the writeup sums up the philosophy: verification makes results "two to three times better", but verification doesn't have to mean a human staring at a diff.
Humans still own the decisions machines are bad at: should we even build this, is the blast radius acceptable, does this pattern set a precedent other engineers will copy, and who's accountable when it ships. Everything else — "is this diff internally consistent with what it claims to do" — got handed to agents that are faster and, it turns out, not obviously worse at catching real bugs.
I'll be honest, the part I didn't expect is the "you land it, you own it" norm surviving at 90% AI-authored code. It works because changes stay small, almost everything is reversible via flags, and the risk tier — not who wrote the diff — decides how much scrutiny it gets. If you're running a smaller team and thinking "we can't do seven layers," you don't need all seven on day one. Start with the risk-tiering idea alone (stop treating every PR the same regardless of what it touches) and you'll get most of the benefit before you've written a line of review automation.
Yes, and the timing isn't a coincidence. On the same day this process writeup went out, Anthropic took Code Review for Claude Code out of preview and made it generally available for Team and Enterprise plans. The product version is the externalized slice of layer 2 and 3 above: it reviews PRs against the full repo, and in GA it's about 3x faster than the preview: most reviews now finish in 6 to 10 minutes instead of roughly 20. Admins can set Quick, Standard, or Deep review depth per repository, which maps directly to the risk-tiering idea from the internal process. Eligible Enterprise orgs can claim at least $10,000 in review credits through January 15, 2027.
So this isn't Anthropic describing a tool they want you to buy and separately bragging about their internal metrics. The product and the process are the same thing, one is just for sale now.
Not the full seven-layer stack on day one. Pick the pieces that transfer without 400 engineers and a dedicated infra team:
CLAUDE.md and a REVIEW.md. Give agents the conventions and review priorities explicitly instead of hoping they infer house style.If you're already running /code-review locally in Claude Code or trying the hosted GitHub product, this writeup is basically the design doc for why it's built the way it is. Worth reading the full thing before you decide how much to trust it on your own repo. We covered the GA launch itself in more detail in our Claude Code Review GA piece, and if you're weighing Claude Code against other agents at this scale, our Claude Opus 5 review covers the model doing the heavy lifting here.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Accenture capped Claude spending at $29/employee/month after some staff ran up $1,000 bills. Here's what that means for Claude Code's weekly limits, and how to stretch your own.
Claude Code Review is generally available for Team and Enterprise, with Quick, Standard, and Deep levels. Here's the setup, real pricing, and the EU credit catch.
Meta's Muse Spark Open is an 816B open-weight model that runs on two Mac Studios. Here's the license, hardware needs, and why it trails Claude Opus 5.5 on coding.